FrontierView to Oxford Economics: 28 AI Representations from the Seller, 4 from the Buyer

A product that works is never interrogated about how it was built. No FrontierView customer has ever asked the company to list, piece by piece, which components of artificial intelligence run inside the service it pays for every month, or what data trained them. On 27 August 2026, someone asked, in writing: whoever was about to buy it.

That day FiscalNote, Inc. — an indirect subsidiary of FiscalNote Holdings, Inc. (OTCID: NOTE, CIK 0001823466) — sold to Oxford Economics USA, Inc., a Pennsylvania corporation, all of the outstanding equity interests in Frontier Strategy Group, LLC, a Delaware limited liability company operating as FrontierView. Signing and closing fall on the same day. The agreement — the Equity Purchase Agreement, filed in full as an exhibit to the Form 8-K the company submitted to the SEC on 2 September (accession 0001193125-26-380190) — commits the seller to dozens of representations it must swear are true. Just one, Section 3.19(o), concerns artificial intelligence: it is where this piece stays.

The inventory nobody asks for until the sale

The section opens with an act no customer in the ordinary course of business ever demands: the Disclosure Schedule must contain “a complete and accurate list of all AI Technology incorporated in any Company Offerings (each, an ‘AI Tool’).” Not a marketing summary: an inventory, piece by piece, attached to a contract rather than a data sheet.

The right to every dataset, one by one

Then comes provenance. The seller swears, verbatim: “Each Acquired Company has complied in all material respects with all licenses, consents, agreements, terms, conditions, written instructions and permissions applicable to the use of each Training Dataset and has a valid and enforceable right to use each Training Dataset as currently used in the operation of its business.” The contract defines the word itself: “training data, validation data, test data, scraped or harvested datasets, or databases, in each case, used to train, finetune, enhance, or improve AI Technology that any Acquired Company uses in the development, training, operation or improvement of any Company Offering.” The definition puts on the record the possibility that some of that data came from scraping, and makes the seller swear that, for each dataset, the right to use it already existed.

What the seller swears it never did

Three denials, not admissions: in each case the seller states that certain conduct did not occur, not that it did. No Acquired Company “has inputted any material confidential information into any AI Technology, except where the provider is contractually prohibited from using or retaining such information other than to provide services to an Acquired Company” — the clause that rules out having fed confidential material into a third party’s model without that contractual guardrail in place. None “has used AI Technology to make recruitment decisions without human review”. And none “has used any AI Technology to generate, create, conceive, reduce to practice or develop any Owned Intellectual Property in any manner that would reasonably be expected to adversely affect, in any material respect, the ownership, validity, enforceability, registrability, or patentability thereof”.

Three years back, and who else would have had to say so

The look-back runs three years: no Acquired Company has received claims that a Training Dataset used in a Company AI Product “was biased, untrustworthy or manipulated in an illegal manner”. A further clause widens the perimeter past outside complaints: “no report, finding or impact assessment of any internal auditor or, to the Knowledge of Seller, external auditor or other Person, makes any such allegation” either. And in those three years no request for information arrived “from regulators or legislators concerning any Company AI Product.” Just before that, the same section requires having “implemented and maintained commercially reasonable processes with respect to its use of AI Technology designed to promote transparency and accountability.” Section 3.20, on personal data, names outright the possibility that an Acquired Company might be acting as a “‘data broker’ as defined under applicable Privacy Laws” — a label the contract writes down only to rule out, not to attach.

Twenty-eight sections against four

The asymmetry is a matter of counting, not rhetoric. In the agreement’s table of contents, Article 3 — the seller’s representations — runs from Section 3.1 to Section 3.28: twenty-eight sections, from corporate organization to patents, from employees to insurance. Article 4 — the buyer’s representations — runs from Section 4.1 to Section 4.4: four, covering essentially organization, absence of violations, and brokers. The buyer promises to pay; the seller swears, section by section, what the company being sold actually is — its artificial intelligence is one of twenty-eight entries in that oath, not a chapter of its own.

The price, with no discovery to announce

The contract defines: “‘Base Purchase Price’ means $7,000,000.” Section 1.2 builds the actual price from there: base, plus the working-capital adjustment, minus debt and transaction expenses, plus the earnout if earned. The 8-K narrative is consistent with that structure, not in conflict with it: it describes “a total value of up to approximately $9.4 million, consisting of $6.4 million in cash at closing and a potential earnout opportunity of up to $3.0 million”. One million dollars of the $6.4 million cash stays “deposited into escrow to satisfy certain potential post-closing purchase price adjustments and indemnification claims”; the earnout depends on “specified annual recurring revenue targets as of each of September 30, 2026 and December 31, 2026.” It should also be said that the 8-K calls these representations “customary for transactions similar to those contemplated by the Purchase Agreement”. It adds that they were made “solely for the benefit of the parties”, to allocate contractual risk between the parties instead of “establishing these matters as facts” — the recurring formula of every M&A 8-K, cited here once as context, not as a discovery.

See the service · Talk to an engineer

What we don’t know

We have not read FrontierView’s list of AI Tools: the Disclosure Schedule referenced in Section 3.19(o)(i) is not filed, and the schedules to an Equity Purchase Agreement are not public as a matter of market practice, not because of an omission specific to this filing. We do not know how many Training Datasets are listed there, or their stated provenance. The revised guidance — 2026 revenue expected between $74 and $76 million, adjusted EBITDA between $8 and $10 million — is stated by the company in its press release, not verified by us against future financial statements that do not yet exist. Nor, finally, do we know whether other recent transactions involving AI companies have included comparable clauses: we have not searched for them, and we do not treat the absence of a search as an answer.

The two axes, applied to this case

Comply. Answering a section like 3.19(o) within a few weeks of a term sheet — the list of AI Tools, the provenance of every Training Dataset, three years of claims and impact assessments with no exceptions — requires a register that already exists, not one assembled during due diligence: for each AI system in production, which data trained it, under what licence, who approved it, and which claims arrived and when.

Decide. The same system holds together models, datasets, contracts and complaints in a single operating model, on which AI agents run the checks and a human operator signs off on the decision. It is the logic of Section 3.19(o), read in reverse: the inventory of what sits inside a product and the right to every dataset are not a box to tick once, at the sale — they are the condition for the value built inside a company to stay inside it, instead of ending up poured into someone else’s model without anyone ever having written it down.

Knowing today how your company would answer a Section 3.19(o) does not require a buyer already at the table. Talk to one of our engineers about it: from the first session, at no cost, comes the dated list of the AI systems you have in production and the datasets that train them.